← Frontier Inference Margins · all research reports
Evidence-scent label pass. Page-set scenario values lose every remaining trace of false empirical scent: "top-decile"/"near-top" stack wording (implying a measured distribution that does not exist) became "aggressive"/"above-baseline" with explicit page-set qualifiers; the utilization slider's "typical" tick — contradicted by its own tooltip's "no representative industry distribution is public" — became "central scenario"; "300B is our median read" became "this page's working estimate"; the "DeepSeek 5:1" tick was corrected to the file's own measured 4:1 profile; "FP8 (frontier default)" is now "(assumed frontier norm)" (closed labs don't disclose serving precision); "evidence median" — naming a fallback that is actually the central scenario, in three places including report prose — now says what the code does; and dossier rows where page-chosen levels wore COMMUNITY ESTIMATE labels were reconciled to SPECULATION. A pre-existing display incoherence was fixed structurally: Kimi/Grok model dossiers annotated traffic parameters with dive-mix sources while rendering Reference-convention values beside them — the dive-only annotations are removed and the dive mixes keep their provenance in the effective-traffic line, where those values actually render. The exploration-surface lint's regex gap (caught "assumption" but not "assumed") is closed. Four GPT-5.6 Sol find→fix→verify rounds, looped to clean on P0/P1.
GPT-5.6 Pro outside-perspective review, adjudicated and remediated. A cold outside-expert review of the live site (18 findings; archived with the adjudication in this annex) was adjudicated finding-by-finding against what the site already says: most findings were already mitigated by welded caveats the reviewer had no context for, several relitigated deliberate design, and six were valid — all cured here. §3's "from first principles" (a derivational status the very next paragraph disowns) became "from an explicit cost identity". §5's "for any frontier model" universal was falsified by the site's own engine (DeepSeek V4 Pro at its own list never enters the 90–95% zone on a GB300 fleet at any utilization) and is now bounded to Western-lab list tariffs with that counterexample stated. The §3/methods-note validity domain now states the size axis plainly: all published serving anchors are ≤~50B-active models while flagship scenarios assume 120–300B active — the model's largest unanchored extrapolation. §5's "margin machine" carries the billable-share caveat inline (86.2%→61.5%). The Zhipu card no longer converts the audited 18.9% segment margin into a "floor" (it is a segment-margin anchor) and its reseller price ceiling is now stated as conditional — the same conditionality the Kimi card always carried. Two currency errors fixed against primary sources: §5's "~500k Trainium" contradicted the site's own hardware table and Anthropic's Apr 20, 2026 statement (">1M Trainium2 in use"); TPU7x GA is Mar 31, 2026 per Google's release notes, not Apr 22 (a Cloud Next conflation).
Provenance upgraded, not just defended. The Anthropic inference-infrastructure margin record (38% → >70%, SemiAnalysis) had been carried as reported-unverified via X relays under the registry's mis-attribution rule ("the newsletter is paywalled"). The margin sentence in fact sits in the newsletter's free portion; it was fetched directly, archived verbatim in this annex (curly apostrophe and all), and the record now quotes the primary source — the same upgrade pattern the Patel/Sequoia citation still awaits. The grounding ledger's version stamp — which claimed "cannot drift" while stamping the previous engine version — regenerates with this release (the no-drift guarantee now includes its own stamp).
Still owed (unchanged). The five-human comprehension test remains the open ship gate; the Patel primary-citation swap and Zephyr re-capture remain flagged owner actions.
The redesign. Person-named analyst presets (TeorTaxes, Zephyr, Dylan Patel/"semi", "skeptic") are reworked into a mechanism-first calculator — the hero margin is still driven only by model/traffic/lens, never by a target — plus a first-class secondary margin-range evidence board fronted by a route-first entry ("what would have to be true for a ~90% margin?") rather than an evidence catalog a reader has to read past. A typed MARGIN_CLAIMS registry bins claims, not people: floors like Dylan Patel's "north of 80%" render as compatible-with every bucket at or above 80 (never as interval membership — a floor is not a ceiling); Zephyr's 90–95% is carried in its own unnamed-subject group, never presented as an Anthropic claimant; company-GM figures (Zephyr's own 70%), reported figures, DeepSeek's disclosure anchor, and the GPT-5.6 Pro model-generated consult each sit in visibly distinct, differently-labeled groups so no serving-context number can pose as the calculator's unit metric. The four analyst presets are absorbed into de-named, page-authored "[range exploration]" routes — numeric-identical to the retired presets, migrated so old permalinks and browser-saved presets resolve to the same numbers under the new ids. Routes exist only where popular discourse actually places a claim: the 60–80% bucket has none (no unit-serving claimant sits there — the central scenario is its own anchor), and the ≥90% bucket has exactly one, because the engine's own reconstruction of the cited "increase the batch size … drive margins from 90% to 95%" mechanism tops out around 87–89% at market rates — the routes that did clear 90% were owned-fleet procurement stories, not the batch mechanism, and were kept separate rather than mislabeled.
Gated the way v2.1.x releases are. A pre-implementation plan review (GPT-5.6 Pro plus a four-persona council) ran before any code, and a second, blind four-persona final gate ran against the built M1–M4 milestones. The final gate returned NO-SHIP, unanimous, with a P0 cluster — all remediated in this release: share-link identity integrity is now fail-closed — a v4 token whose declared model or perspective does not resolve is refused whole and the default state renders, rather than applying the numeric diff under a substituted or mislabeled identity (the exact corruption class v2.1.1 had already fixed once had regressed on two new paths: copying a link after mutating a loaded route, and a crafted token naming a nonexistent model while real selectors stayed on screen); saved presets now carry typed origin and reload as an explicit modified state rather than silently inheriting whatever lens happens to be selected at load time. Board framing lost its "who claims/who places frontier inference margins there" language — an umbrella phrasing that quietly turned a floor into an endorsement despite the correct row-level "compatible with" badge — for explicit relation-aware wording, and reported figures are now labeled "reported figure," never "verbatim," when no archived quotation exists. The hero's ">90%" caption is now interval-aware: it asserts "within the cited 90–95% range" only when the value actually lands there, and says the value sits above that interval otherwise (it had been claiming every ≥90% result, including ~100%, was "within 90–95%"). The model-sizing name block (which had been leaking real names into a loaded route's dossier under a prose disclaimer) moved into a structurally separate, separately-collapsed panel outside the route dossier — quarantine by structure, not by caveat. The evidence board also picked up an intentional 2×2 layout at desktop widths (auto-fit had been stranding the fourth bucket card 3-plus-1) and the v4 permalink schema now validates metadata-carried traffic against the same bounds the sliders enforce, closing an out-of-range/bad-mode injection path.
Also fixed a pre-existing v2.1.2 defect the deep review surfaced along the way: moving a traffic slider or tick updated the hero margin correctly but left the dossier's "Effective traffic mix" line — and the accompanying status note — showing the previous value, because renderAll/fullRefresh never re-rendered the dossier on a plain state change (only certain load paths did). The hero margin itself was never wrong; only its own explanation of itself was stale. The dossier now re-renders on every state change, so the displayed traffic identity always equals the resolved traffic and the number the margin was actually computed from.
Source provenance (honest). The four load-bearing X posts behind the routes and claim records were re-checked against primary sources rather than re-trusted from the archived sweep alone. Dylan Patel's "north of 80%" is now independently confirmed at a primary source — the Sequoia Capital podcast, 2026-07-02, transcribed by three separate outlets — where the site had only a secondary clip account (PodcastAlphaX); swapping the citation to the primary transcript is a recommended follow-up, not yet done. The two TeorTaxes posts and the Zephyr post that the ≥90%/80–90% routes and the Zephyr unnamed-subject claim rest on are unretrievable by automated means — X is login-walled for this research and no Wayback or mirror capture was found — so their durable provenance in this repo is the dated archived research sweep (research/grok-sweep-margin-claims.md), transcribed at the time, not a live-primary re-verification, and that is stated plainly rather than papered over. Zephyr's load-bearing "names no lab" scope point in particular rests on that sweep transcription and is flagged for manual re-capture; the site errs conservative in the meantime — it treats the 90–95% figure as about an unnamed provider and never attributes it to Anthropic.
Still owed (not softened). The five-human, 90-second, 4/5-comprehension-bar test remains open and unrun — it is the actual gate this whole redesign exists to pass, and the two review rounds (plan review, final gate) are a code/copy/provenance pre-check with no established error rate against real readers; they do not substitute for it. A smaller, label-only gap also remains: some status notes omit the traffic-mix figure even though the dossier's own traffic line always carries the correct number (no display shows a wrong value, only an incomplete one in a few paths) — logged in BACKLOG rather than left undocumented.
The plan for this release was itself reviewed pre-implementation by GPT-5.6 Pro and a four-persona council (both archived in this annex); both said "revise before implementing," and the revised plan adopted every P0 condition (the final gate later found one adoption — the traffic state contract's application layer — incomplete, and a second remediation round closed it; see below) — including renaming the new axis honestly (TRAFFIC MIX, not "workload"), spec-first fixtures, and redesigning the planned named-person reception simulations into corpus-bounded source-faithfulness audits with no role-play and no approval outputs.
Traffic mix — I/O + cache (the scoped ontology step; the full resolver remains future work). A third selector owning only ioRatio/cacheHit: provenance-named profiles (no generic "chat/coding" archetypes), model defaults that visibly re-resolve, replays as atomic composites with locked traffic, an anti-lens-shopping span computed at byte-identical traffic ("cost-lens span at [profile]"), v3 permalinks carrying traffic identity, numeric-identity migration for v2 links and browser-saved presets, and a 180-pair frozen baseline proving parity (documented intentional deltas only). The state contract was committed as failing fixtures before implementation.
Coherence cuts. §6 became an assumption-by-assumption divergence table (the full 54-minute consult is now published verbatim with a SHA-256 stamp — recovered from the receiving session's transcript after an earlier revision mislabeled a 39-line digest as "verbatim"); §8 states a longitudinal selection rule; §10 cards carry metric/lens/traffic lines, top-3 evidence lists with the dominant driver marked, and would-update-on triggers.
Durable exports. The full roofline derivation (verbatim, SHA-256), the adopted grounding ledger (machine-generated from the preset registry — the authoritative parameter record), and a dated GPT-5.6 Pro re-emission of the expired preset pack with a checked delta table (the ledger wins).
Final gate — NO-SHIP, then round 2. The four-persona council final gate NO-SHIPPED the first remediation with a precise list: a schema-downgrade permalink forgery (a v3. token declaring an inner v2 schema routed through the lock-overriding legacy-migration path and rendered ≈71.6% under the locked replay's name), type-only sanitizer checks, replay "exit" that was actually a warning on one input path, unimplemented v2/browser-preset identity migration, an incomplete metric rename, a lens span computed at the hidden model default under locked replays, and a release stamp that made the served tree unreproducible. All were fixed in a second round, each encoded as a permanent regression test; the reception resolution ledger's affected rows carry the reopened-and-re-closed history. GPT-5.6 Pro's independent final gate ran in parallel (archived in this annex).
Reception phase — and what it caught. Nine simulated-reader audits ran against the frozen preview (five third-person source-faithfulness audits for TeorTaxes, Zephyr, Dylan Patel, fleetingbits and the @_xjdr deployment numbers; four anonymous archetypes: ML-infra engineer, calibration reviewer, cost-accounting reader, information-scent lurker). They found 16 P0-class defects the two prior review rounds missed — among them: the metric name "list-price" contradicting the realized-billings denominator; a forged-permalink path that bypassed replay-traffic locking; the §3 equation prose claiming interactivity applies to prefill; the TeorTaxes preset pricing H100 below its own cited $2.40 anchor; a site-authored 8:1 ratio presented as ncode deployment telemetry; §1 and dossier scope-expansions of Zephyr's unnamed-lab allegation; unknown /research/ URLs serving the homepage as HTTP 200; and the normalized table falsely claiming the §10 ranges "still apply." All findings are dispositioned in the published resolution ledger (45 fixed after two rounds, 0 rejected — five first-round closings were reopened by the final gate and re-closed; the ledger keeps that history); the hero was de-named (bands describe cited ranges, never people), analyst presets carry always-visible PAGE-AUTHORED RECONSTRUCTION boundaries, precision dropped to whole points with the unrounded value as a diagnostic, and the release gates grew to 499 engine + 220 contract assertions plus a 17-check application-level browser suite that renders the page headless and replays the live attack payloads. Raw outputs, prompts and the corpus manifest are published verbatim. This exercise is not validation and does not satisfy the still-open 5-human-reader test.
Two attribution-fidelity defects found live by the v2.1.2 plan-review round (four-persona council, empiricist) and fixed: the §2c _xjdr blockquote had silently normalized "2.1 sec TTFT overage" to "average" and dropped the "61 sec p95 TTFT (1M ctx)" and "0 chat logs kept" lines — it now matches the archived sweep verbatim ("overage [sic]"), and the prose carries the 61-second p95 as the observed cold-prefill tail. Also repaired a grammatical break the Jukan removal introduced in §6. Both plan reviews are published in this annex (consultation-2026-07-11-plan-review-*).
Removed the §1 "attribution check" passage about Jukan and the related §6/method-note handle-dispute sentences (owner call): since no margin claim of his was ever found, the two research engines' disagreement over his handle changed nothing in the analysis and spent reader attention on research-process trivia. His two substantive citations remain as bare links — "inference is memory" (§4) and the $5B-spend→$15B-ARR lease-economics read (now in §7, labeled a spend-to-revenue multiple, not a margin).
A second four-persona council review plus a GPT-5.6 Pro final gate ran against the deployed v2.1 and returned NO-SHIP with a surgical fix list (archived in this annex). All gate items repaired; the numeric engine is untouched (the reviews verified 5/6 dive replays to a point and called the arithmetic sound):
Backlog implementation round, consulted and gated: a four-persona council design review (typed-ontology and ship-list adjudications), a GPT-5.6 Pro preset grounding pack (first-party tariff verification), and a GPT-5.6 Pro roofline consultation (negative result, adopted) — all archived in this annex.
Position dossiers. Every model and perspective preset now expands into a dossier: whose position it is (quote-anchored), an attribution-honesty line (quoted-position vs reconstruction vs calculator-synthesis — a quote anchors a claim, not a parameter vector), every parameter it pins with source + evidence label, what it assumes away, and what would falsify it. Values render live from the presets (set-canonical), so dossiers cannot drift; the test suite asserts bidirectional key coverage.
Typed scenario ontology. The perspective dropdown now types its entries — [lens] / [analyst] / [replay] / [SLO replay] — making visible that cost-accounting lenses, quoted analyst positions and measured-operating-point replays are different kinds of object. The hero tile shows the margin range across all lenses for the selected model, as a guard against lens-shopping.
New presets (per the council's ship list). Models: DeepSeek V4-Flash (disclosed 284B/13B), GLM-4.7 (disclosed 355B/32B — a distinct price floor from 5.2), and GPT-5.6 Terra/Luna + Gemini 3.5 Flash as clearly-marked TARIFF SCENARIOS (price is the only identified quantity; excluded from the normalized table). Perspectives: xAI cash-marginal and opportunity-cost lenses (xAI-scoped, pairing-warned), China public-cloud on-demand (≈6.15× IDC rates, dated), and an Ant Group H20 SLO replay (Pro tier, measured-operating-point convention). Rejected with reasons: Kimi K2.6 (economically redundant), Grok 4.3 (adds surface, not identification), Anthropic-strategic (merged into the existing strategic-partner lens), Epoch (deferred until a roofline exists), reported-margin-implied (an inverse manifold, not a preset — now a §7 diagnostic sentence). All tariffs verified at first-party sources on 2026-07-10.
Roofline: a documented negative result. The commissioned compute/HBM/fabric roofline (one shared residual, no per-platform efficiency) predicts throughput-oriented operating points within 2–16% but fails the whole-platform gate (Ascend@15ms −39.8%) and the preregistered validation is formally unrunnable on public data (one fitted parameter vs one training platform; no untouched holdout). Anchor fits remain primary; the roofline ships as tests/roofline-diagnostic.mjs and the methods note carries the full result.
Also: permalinkable scenarios (versioned, dated, with a share-time privacy note), pairing warnings for cross-scoped model/perspective combinations, cache-WRITE billing (revenue-side, conservation-safe, default-off; storage remains a documented omission), and the §7 bridge re-framed in dollars per Mtok with a named unreconciled-accounting-perimeter residual.
Revision driven by two independent external reviews, both archived unedited in this annex: a four-persona GPT-5.6 Sol council with Claude Opus synthesis, and a GPT-5.6 Pro adversarial review. Both reached the same verdict — sound thesis, publication-blocking presentation — and their findings substantially overlapped.
Calculator (engine):
tests/snapshots.test.mjs) runs before deploys.Report:
Initial publication: Anthropic-first investigation, interactive calculator, per-provider deep dives (§10), Chinese accelerator support, research annex.